I know from experience that the kernel may respond to ping requests though it could be flat dead for any other purposes. While thinking about it it stroked me that on some cases ping won't even hit the kernel but will be handled by the NIC. And indeed this is the case according to deepseek (below). I guess that a ping to VMs running on hardware equipped with smart NIC the availability test is useless:
The Network Interface (NIC/SmartNIC): Limited Offloading
This is where your question gets interesting. Modern high-performance network interface cards (NICs), especially "smart NICs" or "DPUs" (Data Processing Units), can implement parts of the ICMP logic directly on their own hardware or firmware. This is done for performance reasons, to reduce load on the main CPU.
This is not a full implementation but rather offloading specific, simple tasks:
· ICMP Echo Reply Offload (Ping Offload): This is the most common example. The NIC is configured to listen for incoming ICMP Echo Request packets destined for its IP address. Instead of passing these requests up to the host kernel, the NIC's firmware immediately generates and sends the corresponding ICMP Echo Reply itself. The host CPU is never interrupted, which is a significant performance benefit on servers under heavy load.
· ICMP Rate Limiting: A smart NIC might be programmed to rate-limit certain types of ICMP messages (e.g., Destination Unreachable) directly on the card before they ever hit the kernel, helping to mitigate flood attacks.